CN114126701A - Aortic internal pressure prediction - Google Patents

Abstract

Aspects of the present disclosure describe systems and methods for predicting intra-aortic pressure in a patient receiving hemodynamic support from a transvalvular microaxial heart pump. In some embodiments, the intra-aortic pressure time series is derived from measurements of the pressure sensor of the transvalvular microaxial heart pump, and the motor velocity time series is derived from the measured back EMF of the motor of the transvalvular microaxial heart pump. Additionally, in some embodiments, machine learning algorithms, such as deep learning, are applied to the intra-aortic pressure and motor velocity time series to accurately predict the patient's intra-aortic pressure. In some embodiments, the forecast is short-term (eg, about 5 minutes in advance).

Description

Aortic internal pressure prediction

Cross Reference to Related Applications

This application claims the benefit of U.S. provisional application No. 62/855,389 filed on 31/5/2019, which is incorporated herein by reference.

Technical Field

The present technology relates to systems and methods for predicting intra-aortic pressure (intra-aortic arch pressure) in a patient receiving hemodynamic support from a valvular micro-axial heart pump.

Background

Machine learning has been successfully applied in various technical fields, such as computer vision, natural language processing, speech recognition and clinical healthcare to provide predictions. Examples of machine learning algorithms include Bayesian algorithms (Bayesian algorithms), clustering algorithms (clustering algorithms), decision tree algorithms (decision tree algorithms), dimension reduction algorithms (dimensional reduction algorithms), instance-based algorithms (instance-based algorithms), deep learning algorithms (deep learning algorithms), regression algorithms (regression algorithms), regularization algorithms (regularization algorithms), and rule-based machine learning algorithms (rule-based machine learning algorithms). In clinical healthcare, machine learning algorithms have been used to model risk of death (risk of death), predict hospital stays, detect physiological decline, and classify the phenotype. See, for example, Harutyunyan et al, Multitask learning and benchmark with a clinical time series Data, Scientific Data, doi:10.1038/s 41597-019-; purushothama et al, bearing marking removing models on large healthcare dates, Journal of biological information 83,112, 134, 2018. However, there remains a need for systems and methods for predicting physiological responses that can help physicians to early detect disease and patient response to therapy in real time.

Disclosure of Invention

Heretofore, machine learning algorithms have not been used to predict the aortic pressure (e.g., current aortic pressure, average aortic pressure, median aortic pressure, maximum aortic pressure, minimum aortic pressure, intra-aortic pressure range, systolic aortic pressure, diastolic aortic pressure, etc.) in patients receiving hemodynamic support. Predicting the intra-aortic pressure in a patient is challenging, in part because high frequency intra-aortic blood pressure time series are not currently publicly available. Furthermore, the intra-aortic blood pressure time series may be noisy (noise) and highly unstable. Furthermore, the prediction error and uncertainty of long-term prediction increases dramatically.

The ability to predict the intra-aortic pressure of a patient will greatly enhance the clinician's ability to predict the condition of the patient. For example, Acute Decompensated Heart Failure (ADHF) is a complex clinical event associated with excessive morbidity and mortality, often manifested as a rapid drop in blood pressure with an increase in heart rate. The challenge with ADHF is the lack of effective treatments that both alleviate symptoms and improve clinical outcomes. Existing guideline recommendations are based primarily on expert opinions. See, for example, Givertz et al, ace refined Heart Failure, Update on New and understanding Evidence and orientation for Future Research, Journal of Cardiac Failure, Vol.19, No.6,2013. Therefore, being able to predict the trajectory of the patient's intra-aortic pressure will make it easier for the medical practitioner to assess the patient's risk of ADHF and intervene before collapse (collapse). Furthermore, intra-aortic pressure prediction will provide useful guidance for patients to stop (bathing off) support when their health improves.

Aspects of the present disclosure describe systems and methods for predicting intra-aortic pressure in a patient receiving hemodynamic support from a valvular microaxial heart pump. In some embodiments, the intra-aortic pressure time-series is derived from measurements of a pressure sensor of the trans-valvular micro-axial heart pump and the motor speed time-series is derived from measured back electromotive force (EMF) of a motor of the trans-valvular micro-axial heart pump. Further, in some embodiments, a machine learning algorithm, such as deep learning, is applied to the intra-aortic pressure and motor speed time series to accurately predict the intra-aortic pressure of the patient. In some embodiments, the prediction is short-term (e.g., about 5 minutes in advance).

One aspect of the present disclosure is directed to a system comprising a trans-valve micro-axial heart pump and one or more processors. The trans-valvular micro-axial heart pump includes a motor and a pressure sensor. The one or more processors are configured to: obtaining a set of intra-aortic pressure measurements corresponding to pressure values measured by the pressure sensor during a time period in which the trans-valvular micro-axial pump is at least partially located in the patient's heart; obtaining a set of motor speed measurements corresponding to a rotational speed of the motor during the time period; predicting an intra-aortic pressure of the patient based on the sets of intra-aortic pressure and motor speed measurements using a trained machine learning model; and automatically adjusting the speed setting of the motor based on the predicted intra-aortic pressure of the patient.

In some embodiments, the one or more processors are further configured to obtain a set of current measurements corresponding to an energy intake (energy intake) of the motor during the time period, and the predicting is further based on the set of current measurements.

In some embodiments, the trans-valvular micro-axial heart pump further comprises a tube, an inlet region having one or more openings through which blood can be drawn into the tube by the motor, and an outlet region having one or more openings through which blood can be expelled from the tube by the motor, and the pressure sensor is coupled to the outlet region (coupled). In some embodiments, the transvalve microaxial cardiac pump further comprises an additional pressure sensor coupled to the inlet region, the one or more processors are further configured to obtain a set of left ventricular pressure measurements corresponding to pressure values measured by the additional pressure sensor during the time period, and the predicting is further based on the set of left ventricular pressure measurements.

In some embodiments, the machine learning model is a deep learning model. In some embodiments, the Deep learning model is an Autoregressive Integrated Moving Average (ARIMA) model, a Deep Neural Network (DNN) model, a recursive Sequence to Sequence (recurse Sequence to Sequence) model, a recursive Sequence to Sequence model with Attention mechanism (recurse Sequence to Sequence model with Attention mechanism), a Transformer model, a time domain Convolutional Neural Network (TCN) model, or a Convolutional Neural tower model (Convolutional Neural tower model). In some embodiments, the deep learning model is a recursive sequence-to-sequence model with Legendre Memory Units (LMUs).

In some embodiments, the machine learning model is trained on a dataset with increasing sequences (ascending sequences), decreasing sequences (descending sequences), and stationary sequences (stationary sequences), wherein each sequence comprises intra-aortic pressure and motor speed measurements. In some embodiments, the sequence is incremental if the measure of aortic internal pressure within the sequence increases above a predetermined threshold, decremental if the measure of aortic internal pressure within the sequence decreases above a predetermined threshold, and stationary if the measure of aortic internal pressure within the sequence does not increase or decrease above a predetermined threshold. In some embodiments, the predetermined threshold is 10 mmHg. In some embodiments, each sequence includes a predetermined number of aortic pressure and motor speed measurements. In some embodiments, each sequence includes real-time (RT) aortic internal pressure and motor speed measurements. In some embodiments, each sequence includes mean time (AT) aortic internal pressure and motor speed measurements.

In some embodiments, the machine learning model is trained on a data set having only increasing and decreasing sequences, wherein each sequence includes intra-aortic pressure and motor speed measurements. In some embodiments, the sequence is incremented if the measure of intra-aortic pressure within the sequence increases above a predetermined threshold, and the sequence is decremented if the measure of intra-aortic pressure within the sequence decreases above a predetermined threshold. In some embodiments, the predetermined threshold is 10 mmHg. In some embodiments, each sequence includes a predetermined number of aortic pressure and motor speed measurements. In some embodiments, each sequence includes real-time (RT) aortic internal pressure and motor speed measurements. In some embodiments, each sequence includes mean time (AT) aortic internal pressure and motor speed measurements.

In some embodiments, automatically adjusting the speed setting of the motor based on the predicted intra-aortic pressure of the patient comprises temporarily increasing the speed setting of the motor when the predicted intra-aortic pressure of the patient is less than the current intra-aortic pressure of the patient by more than a predetermined amount.

Another aspect of the present disclosure is directed to a system comprising a trans-valvular micro-axial heart pump, one or more processors, and a display. The trans-valvular micro-axial heart pump includes a motor and a pressure sensor. The one or more processors are configured to: obtaining a set of intra-aortic pressure measurements corresponding to pressure values measured by the pressure sensor during a time period in which the trans-valvular micro-axial pump is at least partially located in the patient's heart; obtaining a set of motor speed measurements corresponding to a rotational speed of the motor during the time period; and predicting the intra-aortic pressure of the patient based on the sets of intra-aortic pressure and the motor speed measurements using a trained machine learning model. The display is configured to display a predicted intra-aortic pressure of the patient.

In some embodiments, the display is configured to simultaneously display the predicted intra-aortic pressure of the patient with the current intra-aortic pressure of the patient and the current speed setting of the motor. In some embodiments, the display is further configured to display an alert when the predicted aortic internal pressure of the patient is less than the current aortic internal pressure of the patient by more than a predetermined amount. In some embodiments, the display is configured to display the predicted intra-aortic pressure of the patient as part of a map.

Yet another aspect of the present disclosure relates to a method for treating a patient with a valvular micro-axial heart pump received into the patient. The method comprises the following steps: inserting a trans-valvular micro-axial heart pump into a patient; obtaining a set of intra-aortic pressure measurements corresponding to pressure values measured by a pressure sensor located on the trans-valve micro-axial heart pump during a time period in which the trans-valve micro-axial pump is at least partially located in the patient's heart; obtaining a set of motor speed measurements corresponding to a rotational speed of the motor during the time period; predicting an intra-aortic pressure of the patient based on the sets of intra-aortic pressure and motor speed measurements using a trained machine learning model; and automatically adjusting the speed setting of the motor based on the predicted intra-aortic pressure of the patient.

In some embodiments, the method further comprises obtaining a set of current measurements corresponding to energy intake of the motor during the time period, and the predicting is further based on the set of current measurements.

In some embodiments, the trans-valvular micro-axial heart pump further comprises a tube, an inlet region having one or more openings through which blood can be drawn into the tube by the motor, and an outlet region having one or more openings through which blood can be expelled from the tube by the motor, and the pressure sensor is coupled with the outlet region. In some embodiments, the transvalve microaxial cardiac pump further comprises an additional pressure sensor coupled to the inlet region, the method further comprises obtaining a set of left ventricular pressure measurements corresponding to pressure values measured by the additional pressure sensor during the time period, and predicting is further based on the set of left ventricular pressure measurements.

In some embodiments, the method further comprises adjusting the amount of the drug provided to the patient based on the predicted intra-aortic pressure. In some embodiments, the method further comprises decreasing the motor speed if an increase in aortic internal pressure is predicted. In some embodiments, the method further comprises increasing the motor speed if a decrease in aortic internal pressure is predicted.

Drawings

Fig. 1(a) illustrates a trans-valvular micro-axial heart pump.

Fig. 1(b) illustrates the trans-valvular micro-axial heart pump of fig. 1(a) positioned within a heart of a patient.

Fig. 1(c) illustrates a ventricular support system.

Fig. 2(a) illustrates information that may be displayed on the main screen.

Fig. 2(b) illustrates information that can be displayed on a placement screen.

Fig. 2(c) illustrates information that may be displayed on a purge screen.

Fig. 2(d) illustrates information that may be displayed on the infusion history screen.

Fig. 2(e) illustrates information that may be displayed on the main screen.

Fig. 2(f) illustrates information that may be displayed on the placement screen.

Fig. 2(g) illustrates information that may be displayed on the placement screen.

Fig. 2(h) illustrates information that can be displayed on the placement screen.

Fig. 3 illustrates a system for monitoring and/or controlling a plurality of medical devices, such as a valvular micro-axial heart pump.

Fig. 4 shows the isovolumic diastole, ejection phase, isovolumic diastole and filling phase of the cardiac cycle.

Fig. 5 illustrates regular waveforms of aortic internal pressure (AoP), Left Ventricular Pressure (LVP), differential pressure (dP), pump flow rate, and motor current, and their relationship to systolic and diastolic.

Fig. 6 illustrates the overall structure of a recursive (recurrents) sequence to sequence model.

Fig. 7 illustrates the overall structure of the transform model.

Fig. 8 illustrates the overall structure of a time-domain convolutional neural network.

FIG. 9 illustrates the overall structure of the time-domain convolutional neural pyramid.

Fig. 10 illustrates a 10 second 25hz (rt) aortic internal pressure, motor speed, and motor current time series.

Fig. 11 illustrates a 20 minute 0.1hz (at) mean intra-aortic pressure time series.

FIG. 12 illustrates the root mean square error of a selected model.

Fig. 13 illustrates MAP prediction for a single recorded real data (ground route) over 24 hours for two deep learning models.

Fig. 14 illustrates MAP prediction for an ascending sequence, a descending sequence, and a stationary sequence for two deep learning models.

FIG. 15 illustrates the root mean square error of a selected model.

Detailed description of the preferred embodiments

Embodiments of the present disclosure are described in detail with reference to the drawings, wherein like reference numerals represent similar or identical elements. It is to be understood that the disclosed embodiments are merely examples of the disclosure, which can be embodied in various forms. Well-known functions or constructions are not described in detail to avoid obscuring the disclosure in unnecessary detail. Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting, but merely as a basis for the claims and as a representative basis for teaching one skilled in the art to variously employ the present disclosure in virtually any appropriately detailed structure.

Efforts have been made to predict the peripheral blood pressure of patients through various machine learning models and statistical methods. See, e.g., Abbasi et al, Long-term Prediction of Blood Pressure Time Series Using Multiple Fuzzy Functions, 21st Iranian Conference on Biomedical Engineering, ICBME, 2014; peng et al, Long-term Blood Pressure Prediction with Deep Current Neural Networks, arXiv: 1705.04524v3, 2018.

Various machine learning models and statistical methods have been used in an effort to predict whether a patient is likely to experience an acute hypotensive Attack (AHE). See, e.g., Henriques & Rocha, Prediction of acid latent impurities Using Neural Network Multi-models, Computers in graphics 36:549552,2009; moody & Lehman, Predicting Acute latent Episodes 10th Annual PhysioNet/Computers in clinical Challenge, computer. Cardiol.,36(5445351) 541-; johnson et al, MIMIC-III, a free accessible crystalline care database, Scientific Data, DOI:10.1038/sdata.2016.35, 2016; hatib et al, Machine-learning Algorithm to Prest Hypotion Based on High-Fidelity Arterial Pressure Analysis, Anesthesiology,129(4), 663-674, 2018.

Efforts have been made to predict Acute Decompensated Heart Failure (ADHF) using various machine learning models and statistical methods. See, e.g., Kenney et al, Early Detection of Heart Failure Using Electronic Health Records, circle, Cardiovasc, Qual, Outcome, 9: 649-; deo & Nallamothu, Learning About About Machine Learning The research and Pitfalls of Big Data and The Electronic Health Record, circ.Cardiovasc.Qual.Outcome, 9: 618-; passantino et al, differentiating mortalities in substrates with access heart failure, Role of risk scales, World J.Cardiol.,7(12):902911,2015; thorvaldsen et al, differentiating skin in Patents Hospitalized for actual compensated Heart Failure and predicted injection Fraction, circle.

However, none of the above cited studies describe a system or method for predicting the intra-aortic pressure in a patient receiving hemodynamic support. Some of the cited studies describe systems or methods for predicting peripheral blood pressure of a patient. However, peripheral blood pressure provides an indirect indication of the patient's cardiac function, while intra-aortic pressure provides a direct indication of the patient's cardiac function. Peripheral blood pressure may be obtained using, for example, a blood pressure tourniquet (cuff) wrapped around a patient's limb (e.g., a cuff tourniquet or a wrist tourniquet), while intra-aortic pressure may be obtained using, for example, a trans-valvular micro-axial heart pump. Thus, peripheral blood pressure provides less information about the patient's condition than intra-aortic pressure.

Furthermore, some of the methods described in the above-cited studies are not practical, at least from a clinical perspective, because they require a large number of input variables. Furthermore, some of the variables used in the above-cited studies are not easily measured. Furthermore, some of the models proposed in the above-cited studies are only suitable for assessing long-term mortality. They do not help physicians to detect diseases, such as ADHF, early in time.

Patients with severe multiple Coronary Artery Disease (CAD), unprotected left main coronary stenosis, last remaining patent vein and/or severely reduced Left Ventricular (LV) Ejection Fraction (EF) are often rejected for cardiac surgery and are increasingly being triaged to (referred for) high-risk percutaneous coronary intervention (FIR-PCI). A valvular microaxial heart pump as shown in fig. 1(a) (e.g., Impella, available from Danfoss, Mass.) Abiomed, Inc., of Mass.)

Are increasingly used during HR-PCI to prevent hemodynamically unstable and improve clinical outcomes. See, for example, Russo et al, Hemodynmics and its modulators transforming a micro-axial-heart-pump-protected PCI in high rise substrates with reduced emission fraction, int.J.Cardiol.274: 221-; dixon et al, A selective reactive three in The using of The transformed Micro-axial Heart Pump system in paper undersurfacing high-risk property in initial U.S. experiment, JACC diov.Interv.2 (2)91-96,2009; o' Neill et al, A productive, random, sequential, vertical, intra-atmospheric, particulate, pump in substrate and underlying, Circulation 126(14) 1717-.

A trans-valvular micro-axial heart pump is a percutaneous catheter-based device that provides hemodynamic support to a patient's heart. As shown in fig. 1(a), the trans-valvular micro-axial heart pump 110 may include a pigtail (pigtail)111, an inlet region 112, a cannula (cannula)113, a pressure sensor 114, an outlet region 115, a motor housing 116, and/or a catheter 117. The pigtail 111 may assist in stabilizing the trans-valvular micro-axial heart pump 110 in the patient's heart. During operation, blood may be drawn into the one or more openings of inlet region 112, directed (channeled) through cannula 113, and expelled through the one or more openings of outlet region 115 by a motor (not shown) disposed in motor housing 116. In some embodiments, pressure sensor 114 may include a flexible membrane integrated into cannula 113. One side of pressure sensor 114 may be exposed to blood pressure outside cannula 113 and the other side may be exposed to blood pressure inside cannula 113. In some such embodiments, pressure sensor 114 may generate an electrical signal proportional to the difference between the pressure outside of cannula 113 and the pressure inside cannula 113. In some embodiments, the differential pressure measured by the pressure sensor 114 can be used to position the trans-valve micro-axial heart pump 110 within the heart of the patient. In some embodiments, pressure sensor 114 is an optical pressure sensor. The catheter 117 may provide one or more fluid and/or electrical connections between the trans-valvular micro-axial heart pump 110 and one or more other devices of the ventricular support system.

As shown in fig. 1(b), the trans-valvular micro-axial heart pump 110 may be positioned in a patient's heart 120. As shown, the trans-valvular micro-axial heart pump 110 may be percutaneously inserted into the ascending aorta 124, through the aortic valve 126, and into the left ventricle 128, e.g., via the femoral artery 122. In other embodiments, the trans-valvular micro-axial heart pump may be inserted percutaneously into the ascending aorta 124, through the aortic valve 126, and into the left ventricle 128, for example, via the axillary artery 123. In other embodiments, the trans-valvular micro-axial heart pump may be inserted directly into the ascending aorta 124, through the aortic valve 126, and into the left ventricle 128, for example. During operation, the trans-valvular micro-axial heart pump 110 carries blood from the left ventricle 128 and discharges the blood into the ascending aorta 124. Thus, the trans-valvular micro-axial heart pump 110 performs some of the work normally done by the patient's heart 120. The hemodynamic effects of the trans-valvular micro-axial heart pump include increased cardiac output, improved coronary blood flow, and thus decreased LV end diastolic pressure, pulmonary capillary wedge pressure, myocardial workload, and oxygen consumption. See, for example, Burkhoff & Naidu, The science after and perfect of culture chemodynamics support a review and composition of support strategies, Catheter Cardiovasc. Interv.80: 816-.

As shown in fig. 1(c), the trans-valvular micro-axial heart pump 110 may be incorporated into the ventricular support system 100. Ventricular support system 100 also includes a controller 130 (e.g., Automated Impella @ from Abiomed, Inc., Danvers, Mass.)

) Display 140, decontamination subsystem 150, connector cable 160, plug 170, and repositioning unit 180. As shown, the controller 130 includes a display 140. The controller 130 monitors and controls the trans-valvular micro-axial heart pump 110. During operation, the purge subsystem 150 delivers purge fluid to the trans-valve micro-axial heart pump 110 through the catheter 117 to prevent blood from entering the motor (not shown) within the motor housing 116. In some embodiments, the purging fluid is a glucose solution (e.g., a 5% aqueous glucose solution containing 25 or 50IU/mL heparin). The connector cable 160 provides an electrical connection between the trans-valvular micro-axial heart pump 110 and the controller 130. Plug 170 connects conduit 117, decontamination subsystem 150, and connector cable 160. In some embodiments, the plug 170 includes a memory for storing operating parameters in the event that the patient needs to be transferred to another controller. The repositioning unit 180 may be used to reposition the valvular micro-axial heart pump 110.

As shown, the purge subsystem 150 includes a reservoir 151, a supply line 152, a purge cartridge 153, a purge tray 154, a purge tubing 155, a check valve 156, an accumulator 157, an infusion filter 158, and a sidearm 159. For example, the container 151 may be a bag or a bottle. The purified fluid is stored in a container 151. A supply line 152 provides a fluid connection between the container 151 and the purge cartridge 153. The purge cartridge 153 may control how purge fluid in the reservoir 151 is delivered to the trans-valve micro-axial heart pump 110. For example, the purge cartridge 153 may include one or more valves for controlling the pressure and/or flow rate of the purge fluid. The purge disk 154 includes one or more pressure and/or flow sensors for measuring the pressure and/or flow rate of the purge fluid. As shown, the controller 130 includes a purge cartridge 153 and a purge tray 154. Purge tubing 155 provides a fluid connection between purge disc 154 and check valve 156. Accumulator 157 provides additional fill volume during purge fluid changes. In some embodiments, accumulator 157 includes a flexible rubber diaphragm (diaphragm) that provides additional fill volume through the expansion chamber. Infusion filter 158 helps prevent bacterial contamination and air from entering conduit 117. A side arm 159 provides a fluid connection between the infusion filter 158 and the plug 170.

During operation, the controller 130 receives measurements from the pressure sensor 114 and the purge tray 154 and controls the motor (not shown) and the purge cartridge 153 within the motor housing 116. As described above, the controller 130 controls and measures the pressure and/or flow rate of the purge fluid through the purge cartridge 153 and the purge disk 154. During operation, after exiting the purge subsystem 150 through the side arm 159, the purge fluid is directed through a purge cavity (not shown) within the conduit 117 and plug 170. A sensor cable (not shown) within conduit 117, connector cable 160, and plug 170 provides an electrical connection between pressure sensor 114 and controller 130. The conduit 117, connector cable 160, and motor cable (not shown) within the plug 170 provide an electrical connection between the motor within the motor housing 116 and the controller 130. During operation, the controller 130 receives measurements from the pressure sensor 114 through the sensor cable and controls the power delivered to the motor within the motor housing 116 through the motor cable. By controlling the power delivered to the motor within the motor housing 116, the controller 130 can control the speed of the motor within the motor housing 116.

Various modifications may be made to ventricular support system 100 and one or more components thereof. For example, as in Abiomed,

the Ventricular Support system 100 may be modified to accommodate other classes of Ventricular Support Systems for Use of Dual geographic Shock and High-Risk PCI as detailed in the Instructions for Use and Clinical Reference Manual, Document No.0042 and 9028 rG (Apr.2020), which is incorporated herein by ReferenceType of valvular micro-axial heart pump, e.g. Impella

Impella

And Impella

A conduit. As another example, one or more sensors may be added to the trans-valvular micro-axial heart pump 100. For example, as described in U.S. patent application No. 16/353,132 entitled "Blood Flow Rate Measurement System," filed on 3, 14, 2019, which is incorporated herein by reference, a signal generator may be added to the trans-valvular micro-axial heart pump 100 to generate a signal indicative of the motor speed within the motor housing 116. As another example, a second pressure sensor may be added to the trans-valvular micro-axial heart pump 100, proximate the inlet region 112, that is configured to measure left ventricular blood pressure. In such embodiments, additional sensor cables may be disposed within conduit 117, connector cable 160, and plug 170 to provide electrical connections between one or more additional sensors and controller 130. As yet another example, one or more components of ventricular support system 100 may be separate. For example, the display 140 may be incorporated into another device that communicates with the controller 130 (e.g., wirelessly or through one or more cables).

Fig. 2(a) - (h) illustrate different screens that may be displayed by the display 140. For example, fig. 2(a) illustrates a graphical representation including a heart pump type 211 (e.g., "Impella 5.0"), a heart pump serial number 212 (e.g., "171000"), a date and time 214 (e.g., "2019-08-2115: 56"), a software version number 216 (e.g., "IC 4048V 8.1"), a power icon 218 (e.g., a battery indicator), button tabs 221, 222, 224, 226, and 228 (e.g., "silent alarm," "flow control," "display," "purge menu," and "menu"), a current heart pump speed (performance) setting 230 (e.g., "P-4"), a heart pump flow measurement 242, a purge system measurement 244, a status indicator 251 (e.g., "Impella Position OK"), 261, and a notification area 270. The current heart pump speed (performance) setting 230 corresponds to the speed at which the motor within the motor housing 116 is operating. For example, "P-4" may indicate that the motor within the motor housing 116 is operating at approximately 22,000 rpm. Cardiac pump flow measurements 242 include the average flow rate (e.g., "1.6L/min"), minimum flow rate (e.g., "1.1L/min"), and maximum flow rate (e.g., "2.1L/min") of blood through the transvalvular mini-axial pump 100. The heart pump flow measurement 242 may be derived from measurements obtained by the pressure sensor 114 and/or the energy intake of the motor within the motor housing 116. The decontamination system measurements 244 include a current flow rate (e.g., "10.2 ml/hr") and a current pressure (e.g., "99 mmHg") of the decontaminating fluid through the decontamination subsystem 150. The purge system measurements 244 may be derived from measurements obtained from the purge tray 154. Diagram 161 illustrates how the trans-valvular micro-axial heart pump 110 should be positioned in the patient's heart. In fig. 2(a), notification area 270 includes notifications 271, 272, and 273.

Each of notifications 271, 272, and 273 includes a header (header) and an instruction set. For example, notification 271 includes the title "Purge System Open" and instruction "1. check Purge System tubing for Open connections or leaks" and "2. press Purge menu soft key, then select Change Cassette & Bag". The notification 272 includes the title "Suction (Suction)" and the instruction "1. lower pressure Level (P-Level)" "2. check fill and volume status" and "3. check Impella position". The notification 273 includes the title "flight mode enabled" and the instruction "1. ground the controller for air transport", "2. enable flight mode on module if Impella connection is provided" and "3. disable flight mode under MENU after arrival at receiving hospital". In other embodiments, the notifications displayed in notification area 270 may have different structures. For example, the title and instructions may be contained in a single box (box) rather than two distinct boxes. As another example, the notification may not include a title. As yet another example, the instructions may be replaced with different types of information, such as an explanatory statement. For example, the notification may serve as an alarm and include a statement describing the reason for the alarm.

Fig. 2(b) - (d) illustrate a placement screen 204, a purge screen 206, and an infusion history screen 208, respectively. The user can toggle between these screens using buttons located next to button tabs 221, 222, 224, 226, and 228. In other embodiments, different user input devices may be used. For example, in some embodiments, the display 140 may be a touch screen and the user may toggle between screens by tapping the button tabs 221, 222, 224, 226, and 228. As another example, in some embodiments, the user may switch between screens using a separate input device, such as a mouse or keyboard.

All data fields from the main screen 202 are included in the placement screen 204, the purge screen 206, and the infusion history screen 208, except for the status indicator 251, the illustration 261, and the notifications 271, 272, and 273. In other embodiments, additional data fields may be added or removed from these screens. For example, in some embodiments, the heart pump type 211 and heart pump serial number 212 may only appear on the main screen 202.

The placement screen 204, the purge screen 206, and the infusion history screen 208 also include additional information. For example, as shown in fig. 2(b), the placement screen 204 includes a placement signal map 252, a placement signal measurement 262, a motor current map 253, and a motor current measurement 263. The position signal plot 252 illustrates pressure values derived from measurements obtained by the pressure sensor 114 over a period of time (e.g., "10 seconds"). The placement signal measurements 262 include an average pressure value (e.g., "9 mmHg"), a minimum pressure value (e.g., "17 mmHg"), and a maximum pressure value (e.g., "76 mmHg") derived from measurements taken over a period of time by the pressure sensor 114. The motor current graph 253 illustrates the current values provided to the motor within the motor housing 116 over a period of time (e.g., "10 seconds"). The motor current measurements 263 include the average current (e.g., "535 mA"), minimum current (e.g., "525 mA"), and maximum current (e.g., "556 mA") provided to the motor within the motor housing 116 over a period of time. In general, the placement signal map 252, the placement signal measurements 262, the motor current map 253, and the motor current measurements 263 can be used to determine the position of the trans-valve micro-axial heart pump 110 within the heart of the patient.

As shown in fig. 2(c), the purge screen 206 additionally includes a purge flow map 254, a purge flow measurement 264, a purge pressure map 255, and a purge pressure measurement 265. The purge flow graph 254 illustrates the flow rate of purge fluid through the purge subsystem 150 over a period of time (e.g., "1 hour"). Purge flow measurement 264 includes the current flow rate of purge fluid through purge subsystem 150 (e.g., "17.9 ml/hr"). The purge pressure map 255 illustrates the pressure of the purge fluid in the purge subsystem 150 over a period of time (e.g., "1 hour"). The purge pressure measurement 265 includes the current pressure of the purge fluid in the purge subsystem 150 (e.g., "559 mmHg"). In general, the purge flow map 254, the purge flow measurements 264, the purge pressure map 255, and the purge pressure measurements 265 may assist in patient management.

As shown in fig. 2(d), the infusion history screen 208 additionally includes an infusion history table 256, glucose infusion measurements 266, and heparin infusion measurements 267. Infusion history table 256 provides a summary of the amounts of purified fluid, heparin, and glucose provided to the patient during each of a plurality of time periods (e.g., "10: 00-11:00," "11: 00-12:00," "12: 00-13:00," "13: 00-14:00," "14: 00-15:00," and "15: 00-15: 08"). The glucose infusion measurement 266 includes the current rate at which glucose is delivered to the patient (e.g., "935 mg/hr"). The heparin infusion measurements 267 include the current rate at which heparin is delivered to the patient (e.g., "935 IU/hr"). In general, infusion history table 256, glucose infusion measurements 266, and heparin infusion measurements 267 may also assist in patient management.

Fig. 2(e) - (h) illustrate how different types of alerts may be presented to a user through the display 140. For example, when the patient has poor natural ventricular function and the controller 130 is unable to determine the position of the trans-valvular micro-axial heart pump 110 within the patient's heart, the main screen 202 may be updated in the manner shown in fig. 2 (e). More specifically, status indicator 251 may be updated to the status "Impella location unknown" and notification 274 may be added to notification area 270. As another example, when the trans-valvular micro-axial heart pump 110 is fully within the patient's ventricle or aorta, the placement screen 204 may be updated in the manner shown in fig. 2 (f). More specifically, notification 275 may be added to notification area 270. As yet another example, when the exit area 115 is located on or near the patient's aortic valve, the placement screen 204 may be updated in the manner shown in fig. 2 (g). More specifically, the notification 276 may be added to the notification area 270. As yet another example, when pressure sensor 114 fails and controller 130 is unable to calculate a heart pump flow measurement 242, placement screen 204 may be updated in the manner shown in FIG. 2 (h). More specifically, the heart pump flow measurement 242 may be replaced by a table of estimated flows and corresponding MAPs and a notification 277 may be added to the notification region 270.

Fig. 3 illustrates a system 300 for monitoring and/or controlling a plurality of medical device controllers, such as controller 130. The system 300 may include medical device controllers 312, 314, 316, and 318, a computer network 322, a Local Area Network (LAN)324, a remote link module 332, a router 334, a wireless access point 336, a cellular site 338, a server 342, a data store 344, an OCR engine 346, and/or monitoring stations 352 and 354. The computer network 322 may include wired and/or wireless segments and/or networks. For example, the computer network 322 may include a wireless network (e.g., a Wireless Local Area Network (WLAN), commonly referred to as "Wi-Fi"), represented by a wireless access point 336, and/or a cellular network, represented by a cellular site 338) that conforms to the IEEE 802.11x standard. As another example, the computer network 322 may include a private and/or public network, such as a LAN 324, a Metropolitan Area Network (MAN), and/or a Wide Area Network (WAN), such as the Internet (not shown).

The system 300 illustrates several different ways in which a medical device controller may be connected to the computer network 322. For example, the medical device controller 312 is directly connected to the computer network 322. As another example, the medical device controller 314 is optionally connected to the computer network 322 through a remote link module 332. As yet another example, the medical device controller 316 is connected to the computer network 322 through the LAN 324 and the router 334. As yet another example, the medical device controller 318 is connected to the computer network 322 through the LAN 324, the router 334, and the wireless access point 336. The medical device controller 318 is also connected to the computer network 322 through a cellular site 338. In other embodiments, medical device controllers may be added to the system 300 and/or removed from the system 300. Further, multiple medical device controllers may be connected to the computer network 322 in a similar manner. For example, multiple medical device controllers may be directly connected to the computer network 322, much like the medical device controller 312.

Server 342 may be configured to request status information from medical device controllers 312, 314, 316, and 318 over computer network 322. In some embodiments, the server 342 automatically and/or repeatedly requests status information. In some embodiments, the status information includes an image of screen content displayed by a display associated with the medical device controller 312, 314, 316, and/or 318. For example, the status information may be similar to the images of any of the screens illustrated in fig. 2(a) - (h). The images may be sent in one or more messages encoded as video frames or sequences of video frames. Further, the video frame(s) may, for example, contain a pixelated copy of the image. In some embodiments, the status information includes information from one or more data fields displayed by a display associated with the medical device controller 312, 314, 316, and/or 318. For example, status information may include information from one or more data fields like heart pump type 211, heart pump serial number 212, date and time 214, current heart pump speed (performance) setting 230, heart pump flow measurement 242, purge system measurement 244, status indicator 251, and/or notification area 270.

The server 342 may also be configured to process received state information. For example, when the server 342 receives an image of screen content displayed by a display associated with the medical device controller 312, 314, 316, and/or 318, the server 342 may parse the image and extract textual information through an Optical Character Recognition (OCR) portion of the image. In some embodiments, the extracted textual information includes information from one or more data fields displayed by a display associated with the medical device controller 312, 314, 316, and/or 318. In some embodiments, server 342 includes an OCR engine for parsing the image and extracting textual information. In some implementations, server 342 communicates with an external OCR engine (such as OCR engine 346) to parse the image and extract the textual information.

The data store 344 may be configured to store unprocessed and/or processed state information. For example, the data store 344 may store images of screen content displayed by displays associated with the medical device controllers 312, 314, 316, and/or 318 and/or textual information extracted from the images by the server 342 and/or the OCR engine 346. The data store 344 may also be configured to provide at least some unprocessed and/or processed state information to the monitoring stations 352 and 354 upon request. The monitoring stations 352 and 354 may be, for example, telephones, tablets, and/or computers. In some embodiments, the monitoring stations 352 and 354 may use cloud-based techniques to securely and remotely display at least some unprocessed and/or processed state information on associated displays. For example, the monitoring stations 352 and 354 may use an online device management system, such as Impella from Abiomed, Inc., Danvers, MA

To securely and remotely display at least some unprocessed and/or processed state information.

In some embodiments, the server 342 and/or monitoring stations 352 and/or 354 may also be configured to remotely transmit instructions to one or more medical device controllers (e.g., medical device controllers 312, 314, 316, and/or 318) within the system 300. For example, if a controller 130 is added to the system 300, the server 342 and/or the monitoring stations 352 and/or 354 may be configured to remotely adjust the power delivered to the motor within the motor housing 116, the flow rate of the purge fluid through the purge subsystem 150, and/or the pressure of the purge fluid in the purge subsystem 150 by remotely sending instructions to the controller 130. In some embodiments, one or more medical device controllers (e.g., medical device controllers 312, 314, 316, and/or 318) within system 300 may offload (offload) one or more computations to server 342 and/or monitoring stations 352 and/or 354. For example, if the controller 130 is added to the system 300, the controller 130 may offload complex calculations (e.g., machine learning algorithms) to the server 342 and/or the monitoring stations 352 and/or 354. To reduce latency, the controller 130 may also offload such computations to another computing device on the same LAN (not shown).

As shown in fig. 4, the cardiac cycle comprises four phases: an isovolumetric contraction phase 410, an ejection phase 420, an isovolumetric relaxation phase 430, and a filling phase 440. During the cardiac cycle, contraction and relaxation of the heart muscle in the heart chamber causes the two valves (mitral valve 452 and aortic valve 454) to open and close due to the pressure difference. During the isovolumetric contraction phase 410, the mitral valve 452 and aortic valve 454 close and the pressure in the chamber 456 increases until it is high enough for the aortic valve 454 to open. During the ejection phase 420, the mitral valve 452 is closed, the aortic valve 454 is open, and blood flows out of the chamber 456 into the aorta. During the isovolumic diastolic phase 430, the mitral valve 452 and aortic valve 454 close and the pressure in chamber 456 decreases until it is low enough for the mitral valve 452 to open. During the filling phase 440, the mitral valve 452 is open, the aortic valve 454 is closed, and blood flows into the chamber 456. The first two phases are called the systolic phase and the last two phases are called the diastolic phase.

Fig. 5 illustrates regular waveforms of aortic internal pressure (AoP), Left Ventricular Pressure (LVP), differential pressure (dP), pump flow rate, and motor current, and their relationship to systolic and diastolic. The AoP waveform corresponds to the pressure in the patient's ascending aorta (e.g., ascending aorta 124). The LVP waveform corresponds to pressure in a left ventricle (e.g., left ventricle 128) of a patient. The dP waveform corresponds to the pressure differential between the patient's ascending aorta and the left ventricle. The pump flow waveform corresponds to the rate at which blood is drawn from the left ventricle into the ascending aorta by a trans-valvular micro-axial heart pump (e.g., trans-valvular micro-axial heart pump 110). The motor current waveform corresponds to the current provided to the motor of the trans-valvular micro-axial heart pump (e.g., the motor within the motor housing 116).

Maintaining a constant average intra-aortic pressure (MAP) is critical to ensure adequate organ perfusion. See, for example, Chemla et al, Mean atmospheric pressure is the geographic Mean of systelic and dienstic atmospheric pressure in restraining humans, Journal of Applied Physiology 99:6,2278, 2284, 2005. Studies have shown that an increased duration of spending a MAP threshold below 65mmHg is associated with a poorer patient prognosis (patient outcomes), such as the risk of death or organ dysfunction. See, for example, Varpula et al, Hemodenamic variables related to outcomeme in peptide shock, Intensive Care Med.31: 1066-; dunser et al, environmental blood pressure reducing early stage and outgome, Intensive Care Med.35: 1225-; dunser et al, Association of intellectual blood pressure and vasopressor load with a social shock motive: a post hoc analysis of a multicenter tertiary, crit. Care Lond. Engl.13: R181,2009. As shown in fig. 5, physiological waveforms obtained using a catheter-based hemodynamic support device (e.g., a trans-valvular micro-axial heart pump) can be a rich source of hemodynamic information. However, there are few warnings regarding the state of the patient based on a predicted time series of MAP using such devices.

Aspects of the present disclosure describe systems and methods for predicting intra-aortic pressure in a patient receiving hemodynamic support from a valvular microaxial heart pump. Early warning of impending intra-aortic pressure changes (e.g., MAP), even if the warning occurs only 5 to 15 minutes in advance, may help prompt management before the patient has a complete hemodynamic collapse. For example, if the intra-aortic pressure of the patient is predicted to increase or remain stable, the clinician may initiate or continue a Percutaneous Coronary Intervention (PCI) procedure. Similarly, if the intra-aortic pressure of the patient is predicted to decrease, the clinician may delay or terminate the PCI procedure. Typically, a significant decrease in the patient's predicted aortic pressure (e.g., a decrease of at least 10mmHg) indicates that the patient's condition is worsening. However, a sustained increase may also indicate that the patient's condition is deteriorating.

Predicting the steady trend of the intra-aortic pressure can also be used as a signal to wean the patient from the trans-valvular micro-axial heart pump. Similarly, the planned intra-aortic pressure can be used to assign a level of support provided to the patient during the evacuation process. For example, a clinician may adjust the pharmacological support provided to the patient based on the predicted intra-aortic pressure (e.g., by adjusting the amount of a drug provided to the patient, such as a vasopressor or inotrope). As another example, the motor speed setting (e.g., current heart pump speed (performance) setting 230) may be manually adjusted by a clinician and/or automatically adjusted by a connected medical device controller (e.g., controller 130) based on the planned intra-aortic pressure. For example, in some embodiments, the medical device controller may be configured to place the patient out of support by automatically and gradually decreasing the motor speed setting over time. In such embodiments, the medical device may, for example, temporarily increase the motor speed setting if the patient's condition is predicted to deteriorate (e.g., the patient's intra-aortic pressure is predicted to significantly decrease).

In some embodiments, a display (e.g., display 140) associated with the trans-valvular micro-axial heart pump may be configured to display the predicted intra-aortic pressure so that the clinician may react accordingly. For example, with respect to the screens illustrated in fig. 2(a) - (h), the predicted intra-aortic pressure may be displayed next to the heart pump flow measurement 242 and/or the purge system measurement 244. As another example, any significant change in aortic pressure (e.g., +/-10mmHg) may result in a notification to be displayed in notification area 270 or an update to status indicator 251. As yet another example, an intra-aortic pressure prediction screen may be displayed that includes a map of predicted intra-aortic pressure over time, much like the placement signal map 252. As yet another example, a graph of predicted intra-aortic pressure over time may be added to the main screen 202, the placement screen 204, the purge screen 206, and/or the infusion history screen 208 and/or replace a data field in one of these screens (e.g., the placement signal graph 252, the motor current graph 253, the purge flow graph 254, and/or the purge pressure graph 255).

As described above, the valvular micro-axial heart pump not only provides hemodynamic support to aid in the restoration of natural heart function, but is also equipped with, for example, one or more sensors (e.g., pressure sensor 114) to capture measurements at origin (origin), rather than at the periphery. In general, measurements obtained from one or more sensors of the trans-valve micro-axial heart pump and operational characteristics of a motor of the trans-valve micro-axial heart pump (e.g., a motor within the motor housing 116) can provide a rich data set to which a machine learning algorithm can be applied to predict the intra-aortic pressure of the patient. For example, a machine learning algorithm may be applied to a set of features including intra-aortic pressure, motor current, motor speed, and/or motor speed settings (e.g., P-0, P-1, P-2, P-3, P-4, P-5, etc., used as an Impella Catheter from Abiomed, Inc., Danvers, MA). The intra-aortic pressure can be derived from measurements obtained from a pressure sensor of a trans-valvular micro-axial heart pump. The motor current may be derived from the energy intake of the motor of the trans-valvular micro-axial heart pump. The motor speed may be derived from measurements obtained by a signal generator of the trans-valvular micro-axial heart pump. The motor speed can also be derived from the back electromotive force (EMF) of the motor of the trans-valvular micro-axial heart pump. In some embodiments, the motor of the trans-valvular micro-axial heart pump comprises three or more motor windings. In such embodiments, the back EMF may be derived from, for example, a measured voltage across the motor windings disconnected from the power supply. In some embodiments, the power source may be in a connected medical device controller (e.g., controller 130).

A variety of different machine learning algorithms, such as bayesian algorithms, clustering algorithms, decision tree algorithms, dimension reduction algorithms, example-based algorithms, deep learning algorithms, regression algorithms, regularization algorithms, and rule-based machine learning algorithms, can be applied to measurements from the trans-valvular micro-axial heart pump to predict the intra-aortic pressure of the patient. Some examples of deep learning algorithms include autoregressive integrated moving average (ARIMA) models, Deep Neural Network (DNN) models, recursive sequence to sequence models with attentions, Transformer models, time domain convolutional neural network (TCN) models, and convolutional neural pyramid models. In some embodiments, these machine learning algorithms may be implemented by a medical device controller (e.g., controller 130) connected to a trans-valvular micro-axial heart pump. In other embodiments, some or all of the processing may be offloaded to another device over a computer network (e.g., server 342).

The ARIMA model is a statistical approach to popular time series predictions. The components of the model are Autoregressive (AR), ensemble, and Moving Average (MA). Thus, the model uses (a) the dependency between the observations and a certain number of lagging observations, (b) the difference of the original observations (subtracting the observations from those of the previous time step) to smooth the time series, and (c) the correlation between the observations and the residual of the moving average model applied to the lagging observations. Additional information on the ARIMA model can be found in Hyndman & Athanasopoulos, Forecasting: principles and practice,2nd edition, Chapter 8ARIMA models, OTexts: Melbourne, Australia, OTexts.com/fpp2,2018, which is incorporated herein by reference.

A feed-forward Deep Neural Network (DNN) may be composed of one input layer, a plurality of hidden layers, and one output layer. DNN can be used in an autoregressive manner. In such embodiments, a single unit can be used in the output layer to construct DNNs to perform one-step-ahead prediction, and to continually recursively feed back predictions for multiple-step-ahead predictions. Additional information on the DNN model can be found in Schmidhuber, Deep Learning in Neural Networks: An Overview, arXiv:1404.7828v4,2014, which is incorporated herein by reference.

Recursive sequence-to-sequence models use an encoder to map an input sequence to a fixed-size vector, and then a decoder to map the vector to a target sequence. Additional information regarding recursive Sequence-to-Sequence models can be found in Sequence to Sequence Learning with Neural Networks, NeurIPS 2014 of Sutskey et al, which is incorporated herein by reference. A Recurrent Neural Network (RNN) model can be used to retain time domain information in time series because its hidden layer can remember the information processed by the shared weights. For the encoder, a bi-directional RNN model may be used so that the model can process data in both the forward and backward directions. In some embodiments, two separate hidden layers may be used and then merged into the same output layer. For the decoder, an RNN model may be used to decode the target sequence from the hidden state. However, due to the disappearance of the gradient, the RNN model has difficulty learning long-term dependencies. Long Short Term Memory (LSTM) cells may alleviate the problem of the gradient disappearance of memory cell states. The overall structure 600 of a recursive sequence-to-sequence model with LSTM units is illustrated in fig. 6. Additional information about LSTM can be found in Hochreiter & Schmidhuber, Long Short-Term Memory, Neural Computation, Volume 9 Issue 8,1997, which is incorporated herein by reference. As used in the remainder of this disclosure, the recursive sequence-to-sequence model with LSTM elements is referred to simply as "LSTM".

Recursive sequence-to-sequence models require that all necessary information input is compressed into a fixed-length vector from which each output time step is decoded. Thus, it may be difficult for the encoder-decoder network to learn all of the useful information. Attention mechanism (Attention mechanism) may be applied to alleviate this problem. Note that the mechanism can learn local information by using the intermediate encoder state of the context vector (context vector) used by the decoder. Thus, in contrast to functions, an attention mechanism may be used to overcome the disadvantages of fixed length context vectors by creating shortcuts (shortcuts) between the context vector and the entire source input. Additional information regarding Attention mechanisms can be found in Luong et al, Effective applications to Attention-based Neural mechanism transformation, arXiv:1508.04025,2015, which is incorporated herein by reference.

Legendre Memory Units (LMUs) further solve the gradient vanishing and explosion problems normally associated with training RNNs by projecting continuous-time signals onto d orthogonal dimensions using a cell structure derived from first principles. The LMU provides theoretical guarantees for learning remote dependence even if the discrete time step deltat is close to zero. This enables the gradient to flow across a continuous history of the internal feature representation. The LMU is a recent innovation, realizes the most advanced storage capacity while ensuring the energy efficiency, and is particularly suitable for the chaotic time series prediction task in the medical field. Additional information about LMUs can be found in Voelker et al, Legendre Memory Units, Continuous-Time retrieval in Current Neural Networks, NeurIPS 2019, which is incorporated herein by reference.

The Transformer model is a transduction model that relies entirely on self-attention (note that attention here is different from that described previously) to compute representations of its inputs and outputs, without the use of sequence aligned RNNs or convolutions. The coding and decoding components are both stacks of the same layers, each layer consisting of two sub-layers: a multi-headed attention layer and a fully connected layer. The decoder has these two layers, but there is an attention layer between them, which helps the decoder to focus on the output of the encoder stack. The Transformer model does not use a single scaled dot product attention, but projects the query Q, key K, and value V to the output, as follows:

the attention function is executed in parallel. In some embodiments, residual concatenation and dropout may be used in the transform model to improve performance. In the context of the present disclosure, since the Transformer model is applied to a digital time series, absolute positions in the input may be used rather than position embedding. The overall structure 700 of the Transformer model is shown in fig. 7. As shown, the encoder includes a multi-headed attention layer and a fully-connected layer, and the decoder includes a masked multi-headed attention layer, a multi-headed attention layer, and a fully-connected layer. Additional information on the Transformer model can be found in Vaswani et al, Attention Is All You Need, arXiv:1706.03762v5,2017, which Is incorporated herein by reference.

The TCN model has a convolution hidden layer and runs on a one-dimensional sequence. Convolutional neural networks create hierarchical representations on input sequences, where nearby input elements (elements) interact at a lower level and distant elements interact at a higher level. This provides a shorter path to capture remote dependencies than the chain structure modeled by a circular network. In some embodiments, the overall structure of the TCN model comprises several volume blocks, followed by one flat layer and several fully connected layers. In some embodiments, to equip the model with a sense of order, the absolute position of the input elements may be embedded. In some embodiments, to avoid the "dead _ relu" problem, a leakage _ relu activation function may be applied to each layer of the TCN model. In some embodiments, dropout may be used to avoid overfitting. In some embodiments, residual concatenation may be used to improve the performance of the TCN model. The overall structure 800 of the TCN model is shown in FIG. 8. As shown, the TCN model includes a plurality of convolutional layers, followed by a flat layer and a plurality of fully-connected layers with residual connections. Additional information about TCN models can be found in Bai et al, An Empirical Evaluation of genetic consistency and Current Networks for Sequence Modeling, arXiv:1803.01271v2,2018, which is incorporated herein by reference.

Advantageously, the TCN model has low memory requirements for training. Table 1 shows the complexity of each layer of the LMU, LSTM, DNN, pyramid, TCN and Transformer models. In table 1, n is the input length, d is the model hidden size, and k is the kernel size. In the case of long sequences, such as 5 minutes of real-time (RT) input sequences (e.g., with 7500 samples), the LSTM model can easily drain all available memory and suffer from gradient vanishing problems. In addition, the Transformer efficiency is very low when the input length is larger than the model hidden size. In contrast, the TCN model can efficiently encode high frequency data.

TABLE 1

In the convolutional neural pyramid model, a series of features are learned in two streams. The first stream across the different pyramid levels enlarges the receptive field. The second stream learns the information in each pyramid level and finally merges them to produce the final result. As shown in FIG. 9, the structure 900 of the convolutional neural pyramid model includes layers from 1 to NLevel, where N is the number of levels. We denote these levels as L_iWhere i ∈ {1, …, N }. Different-scale (differential-scale) contents are coded at each level L_iIn (1). Feature extraction and reconstruction operations are applied to each level separately. L is_iIs down-sampled from L_(i-1)The extracted features. At level L_iAnd 2i convolutional layers are used for feature extraction. The reconstruction operation then fuses the information from two adjacent levels. For example, for L_iAnd L_i+1，L_i+1Is upsampled and then summed with the output from L_iThe outputs of (1) are fused. In some embodiments, the downsampling operation is implemented as a maximum pooling layer (maxporoling layer) and the upsampling operation is implemented as a deconvolution layer. Additional information regarding the Convolutional Neural Pyramid model can be found in Shen et al, volumetric Neural Pyramid for Image Processing, arXiv:1704.02071v1[ cs]2017, which is incorporated herein by reference.

To test the effectiveness of some of the deep learning algorithms described above in predicting intra-aortic pressure, patient data from 67 cases of trans-valvular microaxial cardiac pump was obtained. In 57 of these cases HR-PCI was required (41 were selective, 16 were urgent). The remaining 10 cases were indicated for Acute Myocardial Infarction (AMI) cardiogenic shock (CGS). In addition, another 17 cases of valvular micro-axial cardiac pump were used to compare performance in terms of data volume.

Data from these cases included 25HZ aortic internal pressure, 25HZ motor current, 25HZ motor speed, and other waveforms derived from these three signals (e.g., motor speed setting, left ventricular pressure, and heart rate). The data is captured by a medical device controller (e.g., controller 130) connected to a trans-valve micro-axial heart pump (e.g., trans-valve micro-axial heart pump 110). As used herein, a 25HZ time series is referred to as Real Time (RT) data. Average Time (AT) data is derived from the RT data by averaging every 250 RT data points. In other embodiments, different numbers of RT data points may be averaged together to obtain AT data. In some embodiments, the number of RT data points may be selected based on a desired prediction timescale. Fig. 10 illustrates 10 second samples of a 25HZ RT aortic internal pressure and motor speed time series. Fig. 11 illustrates 20 minute samples of a 0.1HZ AT intra-aortic pressure time series. As shown, the waveform of the average intra-aortic pressure is non-stationary and can indicate long-term trends in intra-aortic pressure and patient condition.

Since features such as motor speed setting, left ventricular pressure and heart rate can be derived from motor speed and intra-aortic pressure, only motor speed and intra-aortic pressure were used to test the effectiveness of some deep learning algorithms as described above. The motor current is also not included as a feature because the averaging sequence contains motor current variations less than the motor speed and the intra-aortic pressure. However, in other embodiments, any of these data sets may be used with or instead of motor speed and/or intra-aortic pressure.

A sliding window was used to generate a sequence of 15,000 samples (10 minutes). Removal of sensor artifacts does not reflect the sequence of physiological MAPs (i.e., less than 50mmHg, greater than 200 mmHg). Intra-aortic pressure changes of greater than 10mmHg are considered significant. These time series fall into three categories: an ascending sequence (I), a descending sequence (D) and a smoothing sequence (S). The overall variation of the ascending sequence and the descending sequence is greater than 10mmHg, and the overall variation of the plateau sequence is less than 10 mmHg. Finally, 50,705 increasing RT sequences, 50,577 decreasing RT sequences, and 419,559 smooth RT sequences were collected. All these sequences are also converted to 0.1HZ AT sequences of length 60.

Ten deep learning algorithms (i.e., ARIMA with mean time (AT) input, DNN with AT input, LMU with AT input, LSTM with attention of AT input, TCN with Real Time (RT) input, Transformer with AT input, pyramid with AT input, and pyramid with RT input) were trained to predict the average aortic internal pressure (MAP) five minutes in advance. In other embodiments, the prediction window may be increased or decreased. For example, in other embodiments, the prediction window may be increased to 10 or 15 minutes. Ten deep learning algorithms were also trained using an RMS-prop optimizer and a learning rate decay of 0.8. Training-validation-test partitioning (split) was used at 60% -20% -20%. Since there are many possible combinations of hyper-parameters, a hyper-parameter random lattice search is performed on 10% of the set of left-out data (hold out dataset). See, for example, Bergstra & Bengio, Random Search for Hyper-Parameter Optimization, Journal of Machine Learning Research 13281, 305, 2012. The hyper-parameter search range can be found in table 2. Root Mean Square Error (RMSE) was used as an evaluation metric. The calculated moving average (moving average) of RMSE on the validation set was used as an early stopping criterion. All tests used the same batch size of 64.

TABLE 2

FIG. 12 provides a comparison of the average RMSE achieved by some tested deep learning algorithms. From left to right, each bar graph provides an average RMSE implemented by an LMU with AT input, an LSTM with attention of AT input, a DNN with AT input, a TCN with AT input, a Transformer with AT input, and a pyramid with AT input. As shown, the model was tested on the incremental (I) data set only, the decremental (D) data set only, the stationary (S) data set only, and the I-D-S data set only. The I-D-S dataset contains all three types of sequences in equal proportions. I. The D and S datasets contained 50,000 sample sequences. The I-D-S dataset contained 150,000 sample sequences. All models were trained on the I-D-S dataset. Overall, the LMU model consistently achieved the best average RMSE score, including an average RMSE of 1.837mmHg over the I-D-S dataset.

Figures 13 and 14 illustrate MAP predictions generated by two best performing models (LMU with AT input, LSTM with attention of AT input). Fig. 13 illustrates a comparison of a single recorded MAP prediction against real data (e.g., real aortic internal pressure) over 24 hours. The black lines are the real data and the colored lines are the model predictions. Fig. 14 illustrates MAP prediction for an ascending sequence, a descending sequence, and a stationary sequence. The dashed line is the real data, while the solid line is the model prediction. The aortic pressure and motor speed for the first five minutes are inputs to generate a predicted aortic pressure value. As shown, both models closely follow the real data.

Table 3 shows all RMSE values (mmHg) for each cohort of models trained on the permutations of the increment-decrement-plateau (I, D, S) data sets. The highest number in each entry is the RMSE result for the combined group. The three values in parentheses are the RMSE on each of the three test sets, containing only the ascending, descending, and plateau sequences, respectively. All results are the average of five runs. The I-D-S training set contains all three types of sequences in equal proportions. Only the I-D training set contains an equal proportion of increasing and decreasing sequences. Only the I-S training set contains an equal proportion of ascending and stationary sequences. Only the D-S training set contains decreasing sequences and smoothing sequences in equal proportions.

FIG. 15 provides a comparison of the average RMSE achieved by some tested deep learning algorithms. From left to right, each bar graph provides an average RMSE implemented by an LMU with AT input, an LSTM with attention of AT input, a DNN with AT input, a TCN with AT input, a Transformer with AT input, and a pyramid with AT input. These models use different training and testing data sets. The light grey portion of each bar represents the predicted performance improvement between the initial patient cohort (N-20) and the current patient cohort (N-67). Each model was trained on the permutations of the increment-decrement-plateau (I, D, S) data sets, as described above with respect to table 3. In addition, each model was tested on only the incremental (I) dataset, only the decremental (D) dataset, only the stationary (S) dataset, and the I-D-S dataset, as described above with respect to FIG. 12. Without stationary sequences in the training set, all models can achieve comparable or even better performance in predicting stationary sequences. Furthermore, the improvements illustrated above for each bar demonstrate the potential for even better model performance as more data is collected in the future.

TABLE 3

In summary, these test results demonstrate that the above-described systems and methods can be used to accurately predict the intra-aortic pressure of a patient. Early warning of the patient of an impending change in aortic pressure, even 5 to 15 minutes ahead of time, can greatly improve clinical outcome. For example, The authors of Wijnberge et al, effective of a Machine Learning-Derived Early Warning System for Integrated hypertension vs Standard card on Depth and Duration of Integrated hypertension Dual Infrared Surgery:. The HYPE Randed Clinical Trial, JAMA, Caring for The Clinical Ill patent, doi:10.1001/jama.2020.0592,2020 observe that The time spent in The event of a Hypotension potential During Surgery is significantly reduced when a clinician is notified using a Machine Learning Warning System. Significant changes in aortic pressure (e.g., +/-10mmHg) can be predicted and the caregiver notified, giving the clinician time to properly intervene before hemodynamic instability occurs. In addition, intra-aortic pressure prediction may assist patients with weaning from mechanical circulatory support after recovery of the natural heart. Predicting MAP ahead of time may also assist in the maintenance/upgrade (evolution) of hemodynamic support, as the level of hemodynamic support may be altered by changing the motor speed of the trans-valvular pump.

From the foregoing and with reference to the various figures, those skilled in the art will appreciate that certain modifications may also be made to the disclosure without departing from its scope. While several embodiments of the disclosure have been illustrated in the accompanying drawings, the disclosure is not intended to be limited thereto, as the scope of the disclosure is as broad in the art as the art will allow and the specification is to be read likewise. Therefore, the above description should not be construed as limiting, but merely as exemplifications of particular embodiments. Those skilled in the art will envision other modifications within the scope and spirit of the claims appended hereto.

Claims (27)

1. A system, comprising:

a trans-valve micro-axial heart pump comprising a motor and a pressure sensor; and one or more processors configured to:

obtaining a set of intra-aortic pressure measurements corresponding to pressure values measured by the pressure sensor during a time period when the trans-valvular micro-axial pump is at least partially in the heart of the patient;

obtaining a set of motor speed measurements corresponding to a rotational speed of the motor during the time period;

predicting an intra-aortic pressure of the patient based on the set of intra-aortic pressure and motor speed measurements using a trained machine learning model; and

automatically adjusting a speed setting of the motor based on the predicted intra-aortic pressure of the patient.

2. The system of claim 1, wherein the one or more processors are further configured to obtain a set of current measurements corresponding to an energy intake of the motor during the time period, and wherein the predicting is further based on the set of current measurements.

3. The system of claim 1, wherein the trans-valvular micro-axial heart pump further comprises a tube; an inlet region having one or more openings through which blood can be drawn into the tube by the motor; and an outlet region having one or more openings through which blood can be expelled from the tube by the motor, and wherein the pressure sensor is coupled to the outlet region.

4. The system of claim 3, wherein the valvular micro-axial heart pump further comprises an additional pressure sensor coupled to the inlet region, wherein the one or more processors are further configured to obtain a set of left ventricular pressure measurements corresponding to pressure values measured by the additional pressure sensor during the time period, and wherein predicting is further based on the set of left ventricular pressure measurements.

5. The system of claim 1, wherein the machine learning model is a deep learning model.

6. The system of claim 5, wherein the deep learning model is an autoregressive integrated moving average (ARIMA) model, a Deep Neural Network (DNN) model, a recursive sequence to sequence model with Attenttion, a Transformer model, a time-domain convolutional neural network (TCN) model, or a convolutional neural pyramid model.

7. The system of claim 5, wherein the deep learning model is a recursive sequence-to-sequence model with Legendre Memory Units (LMUs).

8. The system of claim 1, wherein the machine learning model is trained on a data set comprising an increasing sequence, a decreasing sequence, and a stationary sequence, and wherein each sequence comprises intra-aortic pressure and motor speed measurements.

9. The system according to claim 8, wherein a sequence is incremental if the intra-aortic pressure measurements within the sequence increase beyond a predetermined threshold, wherein the sequence is decremental if the intra-aortic pressure measurements within the sequence decrease beyond a predetermined threshold, and wherein the sequence is stationary if the intra-aortic pressure measurements within the sequence do not increase or decrease beyond a predetermined threshold.

10. The system of claim 9, wherein the predetermined threshold is 10 mmHg.

11. The system of claim 8, wherein each sequence includes a predetermined number of aortic pressure and motor speed measurements.

12. The system of claim 8, wherein each sequence comprises real-time (RT) intra-aortic pressure and motor speed measurements.

13. The system of claim 8, wherein each sequence comprises mean time (AT) intra-aortic pressure and motor speed measurements.

14. The system of claim 1, wherein the machine learning model is trained on a data set comprising only incremental and decremental sequences, and wherein each sequence comprises intra-aortic pressure and motor speed measurements.

15. The system of claim 13, wherein a sequence is incremental if the intra-aortic pressure measurements within the sequence increase beyond a predetermined threshold, and wherein a sequence is decremental if the intra-aortic pressure measurements within the sequence decrease beyond a predetermined threshold.

16. The system of claim 1, wherein automatically adjusting the speed setting of the motor based on the predicted intra-aortic pressure of the patient comprises temporarily increasing the speed setting of the motor when the predicted intra-aortic pressure of the patient is less than the current intra-aortic pressure of the patient by more than a predetermined amount.

17. A system, comprising:

a trans-valve micro-axial heart pump comprising a motor and a pressure sensor; one or more processors configured to:

obtaining a set of intra-aortic pressure measurements corresponding to pressure values measured by the pressure sensor during a time period when the trans-valvular micro-axial pump is at least partially located in a heart of a patient;

obtaining a set of motor speed measurements corresponding to a rotational speed of the motor during the time period; and

predicting an intra-aortic pressure of the patient based on the set of intra-aortic pressure and motor speed measurements using a trained machine learning model; and

a display configured to display the predicted intra-aortic pressure of the patient.

18. The system of claim 17, wherein the display is configured to simultaneously display the predicted intra-aortic pressure of the patient with the current intra-aortic pressure of the patient and the current speed setting of the motor.

19. The system of claim 17, wherein the display is further configured to display an alert when the predicted intra-aortic pressure of the patient is less than the current intra-aortic pressure of the patient by more than a predetermined amount.

20. The system of claim 17, wherein the display is configured to display the predicted intra-aortic pressure of the patient as part of a map.

21. A method of treating a patient with a valved micro-axial heart pump received into the patient, the method comprising:

inserting a trans-valvular micro-axial heart pump into a patient;

obtaining a set of intra-aortic pressure measurements corresponding to pressure values measured by a pressure sensor located on the trans-valvular micro-axial heart pump during a time period when the trans-valvular micro-axial pump is at least partially located in the patient's heart;

obtaining a set of motor speed measurements corresponding to a rotational speed of a motor during the time period;

predicting an intra-aortic pressure of the patient based on the set of intra-aortic pressure and motor speed measurements using a trained machine learning model; and

automatically adjusting a speed setting of the motor based on the predicted intra-aortic pressure of the patient.

22. The method of claim 21, further comprising obtaining a set of current measurements corresponding to an energy intake of the motor during the period of time, and wherein the predicting is further based on the set of current measurements.

23. The method of claim 21, wherein the trans-valvular micro-axial heart pump further comprises a tube; an inlet region having one or more openings through which blood can be drawn into the tube by the motor; and an outlet region having one or more openings through which blood can be expelled from the tube by the motor, and wherein the pressure sensor is coupled to the outlet region.

24. The method of claim 23, wherein the valvular micro-axial heart pump further comprises an additional pressure sensor coupled to the inlet region, wherein the method further comprises obtaining a set of left ventricular pressure measurements corresponding to pressure values measured by the additional pressure sensor during the time period, and wherein predicting is further based on the set of left ventricular pressure measurements.

25. The method of claim 21, further comprising:

adjusting an amount of drug provided to the patient based on the predicted intra-aortic pressure.

26. The method of claim 21, further comprising:

if an increase in the intra-aortic pressure is predicted, the motor speed is reduced.

27. The method of claim 21, further comprising:

if the intra-aortic pressure is predicted to decrease, the motor speed is increased.

Images